The climate system model of the National Center for Atmospheric Research is used to examine the predictability arising from the land surface initialization of seasonal climate ensemble forecasts in current, preindustrial, and projected future settings. Predictability is defined in terms of the model's ability to predict its own interannual variability. Predictability from the land surface in this model is relatively weak compared to estimates from other climate models but has much of the same spatial and temporal structure found in previous studies. Several factors appear to contribute to the weakness, including a low correlation between surface fluxes and subsurface soil moisture, less soil moisture memory (lagged autocorrelation) than other models or observations, and relative insensitivity of the atmospheric boundary layer to surface flux variations. Furthermore, subseasonal cyclical behavior in plant phenology for tropical grasses introduces spurious unrealistic predictability at low latitudes during dry seasons. Despite these shortcomings, intriguing changes in predictability are found. Areas of historical land use change appear to have experienced changes in predictability, particularly where agriculture expanded dramatically into the Great Plains of North America, increasing land-driven predictability there. In a warming future climate, land–atmosphere coupling strength generally increases, but added predictability does not always follow; many other factors modulate land-driven predictability.
There is a large body of literature on the simulation of climate changes, both past and future, with global models. The series of assessments compiled periodically by the Intergovernmental Panel on Climate Change (IPCC) over the last two decades have presented summary reviews and consensus deductions based in part on climate model simulations (Kattenberg et al. 1996; Cubasch et al. 2001; Meehl et al. 2007). Over the same period, a large number of modeling studies have been performed to assess the predictability of the climate system on a range of time scales from intraseasonal to seasonal and longer (for some reviews see Latif et al. 1998, 2006; Shukla and Kinter 2006). However, little work has been published on the characteristics and evolution of short-term (intraseasonal to interannual) climate predictability within a changing climate.
The distinction between initial conditions and forcing or boundary conditions is a key aspect for predictability and derives from the time scales of interest. For intraseasonal to seasonal-scale predictability, which reaches beyond the range of deterministic prediction for the evolution of weather systems, there is a scientific basis to predict anomalies in the mean state and the variability of fluctuations at time scales longer than detailed weather variations (Shukla 1998). Such predictions use models that incorporate a complete global ocean and land surface interacting with the atmosphere. The land, ocean and sea ice states form part of the set of initial conditions along with atmospheric states but hold the key to seasonal predictability. The slowly varying land and ocean states provide both an inertia to the climate system and participate in feedbacks, which may support climate anomalies.
The coupled model used in this study is run with fixed concentrations of atmospheric gases and aerosols, vegetation cover and land use, and configuration of glaciers. Predictability is different from actual skill in prediction, where model forecasts are validated against observations. Predictability can be estimated for a climate model system solely from the variance statistics of the model itself.
Most studies of variability in the context of climate change have focused on decadal time scales that significantly project onto the warming trend (e.g., Hawkins and Sutton 2008; Dunstone et al. 2011) or extreme events in a changing climate (e.g., Weisheimer and Palmer 2005; Kundzewicz et al. 2008), but the link to changing predictability is inferred at best. Even studies of El Niño (e.g., van Oldenborgh et al. 2005; Yeh et al. 2009) focus on the changes in the phenomenon itself and not its predictability. The Ensemble-Based Predictions of Climate Changes and Their Impacts (ENSEMBLES) project explored both seasonal-to-decadal hindcasts and climate change with global and regional models (van der Linden and Mitchell 2009), but little was done to investigate changes in seasonal predictability in a changing climate. Liniger et al. (2007) explored the impact of representing actual recent greenhouse gas trends on the skill of seasonal hindcasts, but the effects on predictability were shown only for near-surface air temperature at 4–6-month lead times for the recent past.
Little more has been done regarding the predictability role played by the land surface in a changing climate. There have been many studies of predictability from the land surface in current climate (e.g., Koster et al. 2000; Schlosser and Milly 2002; Dirmeyer 2003, 2005; Kanae et al. 2006; Conil et al. 2007; Douville 2010; Guo et al. 2012). Seneviratne et al. (2006) used a regional model to examine how a warming climate could affect land–atmosphere coupling over Europe. Soil moisture effects on climate have been reviewed by Seneviratne et al. (2010) in the context of climate change with some implications for predictability.
Dirmeyer et al. (2012) examined changes in a variety of metrics for land–atmosphere interactions in multidecadal simulations of current and future (time slice) climate using the forecast model of the European Centre for Medium-Range Weather Forecasts (ECMWF). That study, along with an examination of changing land–atmosphere interactions in a large suite of integrations from phase 5 of the Coupled Model Intercomparison Project (CMIP5) (Dirmeyer et al. 2013), hinted that predictability associated with land–atmosphere coupling could be changing as climate changes. Although techniques exist for estimating the fraction of total variance in weather and climate variables that is forced by the lower boundaries (Madden 1976; Feng et al. 2011, 2012; DelSole et al. 2013a), these techniques have not been applied to discern changes in such predictability.
To explore how climate predictability may evolve from the past though the present to potential future climate scenarios, we have conducted a series of modeling experiments using the ensemble framework for preindustrial and projected future climate, as well as current conditions, with the same climate model. Section 2 describes the experiment design; the ensemble members are constructed and initialized in such a way to isolate the role of land surface initialization from that of ocean and atmosphere initialization. Section 3 illustrates the models' ability to simulate aspects of the current climate, and its intrinsic predictability (effectively the model's ability to predict itself) is shown in section 4. Before demonstrating how predictability in this model evolves from the past to current to future climate in section 6, we shed light on some likely aberrant behaviors in section 5. Conclusions are given in section 7.
2. Model and experiments
For all experiments, we use version 4 of the Community Climate System Model (CCSM4) of the National Center for Atmospheric Research (NCAR). CCSM4 has been integrated at 0.9° × 1.25° resolution. The atmospheric general circulation model component, version 4 of the Community Atmosphere Model (CAM4), uses a finite volume core (Lin 2004) for dynamical computations and includes a number of changes and improvements over previous versions to parameterizations of deep convection, low clouds, and gravity waves (Gent et al. 2011). CAM4 is coupled to version 2 of the Parallel Ocean Program described by Danabasoglu et al. (2012).
The land surface model in CCSM4 is version 4 of the Community Land Model (CLM4; Lawrence et al. 2011). CLM4 has the option of being integrated either with specified vegetation phenology based on satellite observations or with prognostic phenology from a carbon–nitrogen (CN) biogeochemical model. For these experiments, the CN model is enabled; however, the option to dynamically predict the changing distribution of vegetation is not used. As a result, the mean, annual cycle, and interannual variations of vegetation properties, such as greenness and leaf area index (LAI), are forecast, but the spatial distribution of vegetation, including the percentage coverage of different plant functional types within each grid box, are prescribed.
Three long simulations of 50 yr have been conducted, branching off existing simulations conducted at NCAR. One simulation is a preindustrial climate simulation beginning from year 1201 of a 1300-yr run with a perpetual 1850 atmospheric composition and land cover (cataloged at NCAR as b40.1850.track1.1deg.006). The second simulation is branched from a transient late twentieth-century experiment (b40.20th.track1.1deg.005) at year 2000, after which atmospheric composition and land cover are held at 2000 values. The final is a future climate simulation based on the 8.5 W m−2 representative concentration pathway (RCP8.5) scenario (b40.rcp8_5.1deg.001) run from the 1 January 2101 initial state of an 1850–2105 transient simulation. In this case, the atmospheric composition and land cover are held at the 2095 values for the duration of the simulation.
From each 50-yr baseline simulation, 15 years are chosen for conducting sets of 3-month 14-member ensemble forecasts (the baseline is used as the fifteenth member) starting from 0000 UTC 1 May, 1 June, and 1 July. The years for each start date are chosen to sample a range of ENSO states as a means to ensure the full span of model climatological variability would be sampled. Niño-3.4 index time series are calculated from the CCSM4 output for each baseline run based on the mean annual cycle and anomalies for each run. The years for the 15 seasonal simulations are then chosen by the following method. For a given start month, the 3-month values for the Niño-3.4 index [e.g., March–May (MAM) for the 1 May starts] are calculated for each of the 50 years of the baseline simulation. Five cases are chosen from each of the high-index (El Niño), low-index (La Niña), and neutral years. First, the years are ranked by the Niño-3.4 index from high to low, and the El Niño cases are parsed, starting with the year having the highest index. Years are then chosen by going down the ranked list, skipping any year that is adjacent to a year already selected (no situations with two El Niño years in a row). Once five El Niño cases are selected, five La Niña cases are chosen by stepping up the ranked index list from the bottom, assuring no years are retained that are consecutive with either other La Niña or El Niño years. Finally, the five years with the lowest absolute value of the Niño-3.4 index are chosen as neutral years, without regard for sequence or proximity to strong ENSO years. This procedure is repeated independently for each of the three initial dates.
For each initial date, the 14 additional ensemble members are initialized with either identical atmosphere, sea ice, and ocean states from the baseline simulation or identical initial land states. Land states include soil water and ice content, soil and vegetation temperatures, snow, and a large set of variables involving vegetation phenology. Each ensemble member takes its remaining initial states from each of the other 14 years for which seasonal simulations are conducted. Figure 1a presents a schematic of how such an ensemble is assembled, for the earliest of the 15 cases taken from a 50-yr simulation. In this approach, we can consider either ensembles where each member has identical initial atmosphere, sea ice, and ocean but one initial land state from each of the 15 years or ensembles where members have identical initial land states but the remaining initial conditions are from each of the 15 years.
From this arrangement of experiments, we can distinguish the impact of initialization of each of the three components, provided we assume that the atmosphere and ocean initialization impacts can be separated by their different time scales of influence. The atmospheric initialization should primarily influence the first 1–2 weeks of the ensemble, while the effects of ocean initialization should emerge during that same period and persist throughout the season.
Finally, there is an additional set of simulations branched off of the 50-yr simulations that have identical initial atmosphere, sea ice, and ocean states but land states that are very similar (Fig. 1b gives an example for an ensemble starting on 1 June). This is analogous to classic ensemble weather forecast initialization techniques in that the initial perturbations are confined to one element of the system (here it is land instead of atmosphere) and are small relative to the anomalies in the initial state. The additional 14 ensemble members for each case take their land surface states from the same 50-yr simulation 24, 48, 72, … , 168 h before the start of the forecast. This provides seven additional initial states. Another seven, indicated by the white diamonds in Fig. 1b, are manufactured by averaging the adjacent states at each interval (0 with 24-h previous, 24 with 48-h previous, etc., to 144 with 168-h previous). There are two reasons for this approach. The reason to select states only preceding the initial time for the ensemble, instead of centering on the ensemble start time, is to provide states that do not contain future information and are thus representative of what could be done with real-time forecasts. However, we found that land states begin to diverge significantly beyond more than one week. Thus, we used only the previous week's data to create additional states. The 1200 UTC states would be inconsistent with the local diurnal phase at the start time of the forecasts, particularly for temperature, so the 0000 UTC states are averaged instead.
The key ensemble statistics of grand and ensemble means are given as
where xen denotes the value corresponding initialization for year n = 1, 2, … , N, and e = 1, 2, … , E, indicating the ensemble member for that initial state. In all experiments, the total climatological variance of variable x across forecasts initialized in a particular month is
This total variance is the sum two parts: the variance of ensemble means across the different initial states for the same month of different years (i.e., signal),
and the variance of individual ensemble members about their corresponding ensemble means (i.e., noise),
As shown in Fig. 1, both N and E = 15 in these experiments. Signal-to-total ratios are used to quantify predictability, rather than signal-to-noise ratios, because the formulations above imply the maximum value of VS is VT, so 0 ≤ VS/VT ≤ 1.
3. Model climate
We refer the reader to Straus et al. (2013) for a broad scientific motivation and detailed documentation of the experiments and for the basic model climatologies and to DelSole et al. (2013b) for descriptions of the role of ocean variability in seasonal climate predictability. Here, we briefly illustrate some aspects of the climatology of the model for the current (ca. 2000) climate state that will be useful for interpreting the predictability and its changes in the rest of this study.
Figure 2 shows the mean temperature for May, July, and September from the 50-yr late twentieth-century experiment. Also shown for comparison are the same statistics for the last 50 complete years of the Climate Research Unit observed temperature analysis, version 3 (CRU; Jones and Moberg 2003). We focus on this time of year since most of the 3-month ensembles span the period from May to September. The bottom panels show the interannual standard deviation of monthly means averaged for the months of June, July, and August. Most of the variability is associated with the retreating snow and active storm tracks at high northern latitudes including eastern Europe, but pockets of relatively high variability (σ > 1.5 K) exist over the central United States, parts of southern Europe, the Middle East, North Africa, the Indus, Ganges, and La Plata basins, and central Australia. Variability is weakest in the humid tropics. CCSM4 appears to overestimate the high-latitude variance.
We do not show precipitation variability because we find its predictability to be exceptionally low in this model, as discussed later. However, another component of the surface water cycle (i.e., soil moisture) is the principal land surface state for forecast initialization during this period of the year and strongly influences temperature predictability (cf. Koster et al. 2011). Figure 3 shows the same climatological statistics for volumetric soil wetness in the top six layers of the soil (approximately half a meter in depth). Values from CLM4 when coupled within CCSM4 and run offline with observationally based forcings are shown. We do not show a comparison to observations here because there is no long-term global observational dataset—even remote sensing–based estimates span less than two decades and can provide information for only the top few centimeters of soil at best (e.g., Burke et al. 2003). This model's patterns are consistent with other global analyses and the model's own precipitation climatology. Most of the differences between the two implementations are due to differences in precipitation, stemming from systematic errors in CAM4. Precipitation variability is highest in the tropics, but soil moisture variability peaks in areas outside the tropics where conditions are drier and horizontal gradients tend to form (e.g., in the Northern Hemisphere, the Mississippi basin, northeastern China, northern India, eastern Europe, and the Sahel). A priori, it would be reasonable to expect the greatest predictability from realistic land surface initialization to be in these areas (Koster et al. 2006). As we shall see, this is not the case in this climate model.
4. Model dependence
Before looking at predictability in this climate model, we should note how this models' simulation of land–atmosphere interactions differs from other data.
Figure 4 shows the global distribution of the correlation between daily soil moisture and surface latent heat flux for June. This quantity is shown for the current climate simulation with CCSM4 as well as estimates from several other sources: an offline multiple land model estimate [the Second Global Soil Wetness Project (GSWP-2); Dirmeyer et al. 2006], a reanalysis product [Modern-Era Retrospective Analysis for Research and Applications (MERRA); Reichle et al. 2011], and a global weather model [ECMWF Integrated Forecast System (IFS); Dirmeyer et al. 2012]. For the correlations using surface soil moisture (ostensibly the top model layer in each case, which may range from 2 to 10 cm in thickness), all estimates have very similar distributions. High values are seen in arid and semiarid regions, negative values are over wet or radiation-limited locations, and semihumid regions generally have weak correlations. When correlations with the root zone (in the range of 0.5–1.0 m below the surface, depending on the model) soil moisture are considered, more discrepancy is seen, particularly in terms of the areal coverage of strong positive values. CCSM4 is particularly feeble, suggesting the connection between subsurface soil moisture states and surface fluxes in this model is very weak. The same is seen for other months as well (not shown).
Besides the ability to communicate land surface anomalies to the overlying atmosphere, another key aspect of predictability on subseasonal to seasonal time scales is the ability to maintain anomalies. Again, CCSM4, or specifically CLM4, seems to have too little soil moisture memory. We performed a comparison of the evolution of soil moisture memory, calculated as the lagged autocorrelation of soil moisture in the top 10 cm of the soil from 1 June initial conditions (ICs). Comparison is made to observations from the Soil Climate Analysis Network (SCAN; Seyfried et al. 2005) within ±14 days of the same date. Of course, there is a spatial scaling issue comparing point observational data to gridbox model output. Unfortunately, the SCAN network is not dense enough to aggregate usefully to the climate model grid. With few exceptions, there would be only one or two SCAN sites per model grid box. Thus, we make a direct comparison of the data noting that we would expect more variability and thus weaker correlations for the point observations than gridbox averages. As a basis of comparison we looked at the behavior of the fully coupled CCSM4 and CLM4 run in offline mode driven by the meteorological forcing data of Sheffield et al. (2006), as well as the data from GSWP-2.
Figure 5 synthesizes the results. Gridded data were interpolated to each SCAN site and soil moisture instrument with a sample size of at least 30 for the lag in question, and then autocorrelations were calculated. The 7-day means are compared to the initial 1-day means. The series of 30 circles represent progressive lags from the 1–7-day mean to 30–36-day mean, with the shortest lags at the upper right corner, representing the highest correlations in the SCAN (abscissa) and model (ordinate) data. The horizontal and vertical lines through each circle show the total range of values of the lagged autocorrelations across all sites, and the circles themselves are placed at the mean values across all sites. The X = Y diagonal is shown as a dashed–dotted line.
The plot for CLM4 shows that at all lags it has a tendency for lower autocorrelation than the SCAN data, despite the fact that it represents a value over a large area (model grid boxes across the sites range from 7.4 × 104 to 9.9 × 104 km2 in area). The values from CCSM4 are actually slightly higher than for the offline model, very close to the SCAN values in the mean. In contrast, GSWP-2 shows the expected behavior, with substantially higher values of mean lagged autocorrelation than the SCAN sites at all lags because of the difference in spatial scales. SCAN sites show a larger intersite range of values than any of the models, which is to be expected from the scale differences, and more spatial variability. But there seems to be too little memory of soil moisture anomalies in CLM4 relative to either other land surface models (e.g., GSWP-2) or observations when scale differences are considered.
One suspected cause of the weak memory in CLM4 is the values of hydraulic conductivity, which may be too high (D. Lawrence 2012, personal communication). To test this hypothesis, we reran the offline CLM4 simulation with the values of hydraulic conductivity reduced by a factor of 10 everywhere and repeated the estimates. The impacts on soil moisture memory were minimal, and the hypothesis was ruled out. It should be noted that running CLM4 offline with the CN module (predicted vegetation phenology) disabled (prescribed satellite phenology or SP mode) had little impact on soil moisture memory.
Finally, there is indication that processes in the atmospheric boundary layer (ABL) may mute the response to the land surface. Figure 6 shows metrics of local land–atmosphere coupling during June-August (JJA) derived from Santanello et al. (2009) for CCSM4 and the ECMWF IFS as derived globally in Dirmeyer et al. (2012). The top panels show the implied total heating of the ABL estimated from the diurnal range of temperature under the assumption that the ABL is well mixed, and the daily increase of near-surface potential temperature from morning to sunset is an indication of the average change in ABL potential temperature and thus heat content. The second row shows the estimated contribution from surface sensible heat flux, calculated using Eq. (1) of Santanello et al. (2009). The difference is shown in the third row, which should include ABL heating from other processes (mainly entrainment at the top of the growing boundary layer but also advection), although in this calculation we included all days, not only dry days, so moist thermodynamic processes above the surface could act as heat sources and sinks as well.
The surface heating is similar between the two models, and in fact the patterns and magnitudes of surface sensible heat flux and ABL depth are comparable (not shown). But the total heating in CCSM4 is much larger than in IFS, because of a larger near-surface diurnal temperature cycle. So although the magnitude of ABL heating from the land surface flux is about the same, the relative contribution of the land surface is generally much larger in IFS than CCSM4 (bottom panels). So either CCSM4 demonstrates much stronger ABL heating than IFS from atmospheric processes like entrainment, or the assumption of a well-mixed ABL is not well satisfied in CCSM4 such that near-surface potential temperatures are not a good proxy for ABL states. In either case, the impact of the land surface on the lower troposphere appears to be inhibited in comparison to the IFS.
In summary, Fig. 6 suggests the degree to which boundary layer properties in CAM4 that are affected by surface fluxes versus entrainment (and lateral advection) differ substantially from a well-validated operational forecast model. Of course, there are likely errors in both models, but the large discrepancies are troubling. Meanwhile, Figs. 4 and 5 show subsurface soil moisture memory and its control on surface fluxes are rather weak, implicating the land model (CLM4), although certainly near-surface atmospheric conditions have some effect on possible biases in latent heat fluxes that could affect the relationship. These results suggest CCSM4 may be on the weak side of the spectrum of climate models in terms of the absolute strength of land–atmosphere coupling, and at least for the land surface component this assertion is corroborated by observations. With this in mind, we examine the land surface contribution to subseasonal-to-seasonal predictability in this model.
5. Predictability in the current climate
Signal-to-total variance ratios of near-surface air temperature for current climate conditions are shown for the 1 June initializations in Fig. 7. This quantity represents the intrinsic predictability arising from the initial atmosphere and ocean initialization, where land initial states are randomized (as in Fig. 1a). To bring out the features, 30-day averages at 15-day intervals are shown, starting with days 16–45 after most of the predictability from the atmosphere is lost. The lowest contour interval corresponds to the 95% confidence level based on a transformation of Fisher's F test (Guo et al. 2011).
The largest contiguous region of high predictability at all forecast lead times is over the eastern tropical Pacific Ocean, where ENSO variability dominates. This is also true for 1 May and 1 July initializations. The areas of highest predictability span much of the tropics and subtropics, especially over the ocean, and gradually diminish over time. Many tropical land regions also show high predictability of near-surface air temperature. There is little predictability over land in the extratropics beyond the first month.
There are broad areas of relatively high predictability over India, much of Africa, and northern South America in the 16–45-day period, corresponding to late June and early July, but these areas diminish in size and strength in the latter forecast periods. There are, however, broad areas of the northern Pacific and Atlantic that show high predictability during boreal summer that are not present in boreal winter. Figures for the 1 May and 1 July initialization cases are not shown but are broadly consistent with the 1 June case. Some specific features, such as the existence and dissipation of predictability over tropical and subtropical land areas, are mainly a function of the calendar rather than forecast lead time; that is, predictability persists more within the May case and disappears quickly in the July case. Predictability over the extratropical northern oceans also diminishes quickly in the July initialization case.
Figure 8 shows the signal-to-total ratios as a measure of near-surface temperature predictability from the initialization of the land surface states alone (atmosphere and ocean initial states are scrambled). Again, most of the regions of significant predictability are in the lower latitudes, predominantly over areas that are in the dry phase of their annual monsoon cycle. The reason behind this appears to be model specific and is discussed later in this section. However, in the 1 June initialization case, there are weak but significant areas of predictability near the Arctic in the 16–45-day period, associated with the melting of snowpacks that retain memory of initial snow cover anomalies.
Perhaps more interesting are fainter regions over southern Asia, the eastern Sahel of Africa, and especially the central United States that correspond to traditional “hot spots” of land–atmosphere coupling (Koster et al. 2004). These features are also present in the 1 May and 1 July initialization cases and fade gradually as the forecasts progress. The signals overall are weaker than in Fig. 7, as the initial atmospheric and ocean states are essentially randomized among ensemble members. Thus, local synoptic situations and remote oceanic climate anomalies are not synchronized between ensemble members, erasing ensemble-mean anomalies over land even more quickly than in Fig. 7.
The impact of land initialization is more clearly demonstrated by comparing the signals between the two ensembles that have the ideal situation of identical initial atmosphere and ocean states between members. Figure 9 shows this measure of near-surface air temperature predictability (log of the signal-to-signal ratio) for the cases initialized at the start of May, June, and July. The numerator is the estimated signal from the ensembles with small perturbations in the initial land state (from the preceding 7 days; see Fig. 1b) and the denominator comes from the ensembles where each member has an initial land state from completely different years (Fig. 1a). Results are averaged over the second and third months of the forecasts. The increase in predictability (log of the ratio greater than zero) is due to the consistency in the initial land surface states and is analogous to that found in the second phase of the Global Land–Atmosphere Coupling Experiment (GLACE-2) forecasts (Guo et al. 2011). South America, southern Africa, and northern Australia show prominently, but now much of North America, eastern Europe, and southern and eastern Asia consistently show an increase in land-driven predictability, consistent with so-called hot spots of land–atmosphere coupling found in previous studies (e.g., Koster et al. 2004). Pointwise, the 95% confidence level based on an F test and a sample size of 15 includes only the most extreme colors.
The strong land-induced signal-to-total ratio over the Southern Hemisphere subtropics (namely the Cerrado of Brazil, the African steppes and savannas from Namibia to the Rift Valley, and northern Australia) in Figs. 8 and 9 is due to a spurious periodicity that arises at the start of the dry season in tropical (C4 plant type) grasses. Figure 10 shows a time series of the LAI averaged over the Mato Grosso region of Brazil for grid boxes with at least 67% coverage by the plant functional type representing C4 grasses. The colored lines show the evolution of LAI in the 15 individual ensembles from the 2000 case. This is the season where soil moisture stress triggers a drop in LAI some weeks after the start of the dry season.
In the model, the LAI in any given year tends to drop quickly, but the starting value and the date when the drop begins vary from year to year. However, there is very little variation among the ensemble members, even out to 90 days, compared to the swings in ensemble-mean LAI during the dry season. When the LAI is anomalously low, it tends to reverse sign to very high values after 6–7 weeks. There is also a tendency for positive initial anomalies in LAI to reverse as well, sometimes even more quickly. In both cases, the initial sign of the anomaly tends to restore during the third month. These oscillations have strong effects on surface fluxes, soil moisture, and atmospheric states.
The tendency for periodicity is a direct function of the degree of coverage by C4 grass. When we relax the coverage threshold to 50% or 33%, the curves for each year smooth out (not shown). Finally, we have examined other regions that have large amounts of C4 grass. Southwestern Africa and northern Australia, which also show considerable predictability in Figs. 8 and 9, also show dry-season oscillations during this period. There is also a large fraction of C4 grass coverage across the Sahel, but being in the Northern Hemisphere the oscillations there occur from December to April. These findings were presented to the NCAR Land Model Working Group in February 2013, where it was determined that the periodicity likely results from the stress deciduous algorithm in the C4 grass parameterization that includes several 15-day time accumulators (cf. Oleson et al. 2010).
These oscillations and their impacts on regional climate are likely not a useful form of predictability, as this seems to be a spurious behavior of this plant functional type in CLM4. Thus, the apparent signal in these regions should be disregarded. The apparent weakness elsewhere in land-driven predictability, expressed in the signal-to-total ratio, is diagnosed in the next section.
A global indication of the effect of land surface initialization is shown in Fig. 11. The global (land points between 60°S and 60°N) average of the root-mean-square difference among ensemble members for top 1-m soil wetness, 2-m air temperature, total evapotranspiration, and precipitation are shown as an indication of intraensemble spread. Results for May, June, and July initializations are combined here as a function of time from forecast initialization, and a 5-day running mean is applied to remove noise. The red line is for the ensembles constructed so that initial land states are maximally perturbed, and blue is for the ensembles that have small initial land perturbations derived from the days before forecast initialization. Initial atmosphere and ocean states are the same within each ensemble in both cases. The envelopes around each curve indicate 95% confidence limits.
Two interesting conclusions can be drawn from this figure. First, the differing time scales of the influence of land surface initialization can be discerned. The intraensemble spread of soil moisture remains significantly different between the two cases throughout the 90-day period. However, evapotranspiration over land has a time scale of about 20 days globally before the spread within the two sets of ensembles becomes indistinguishable. For near-surface air temperature the time scale is about 16 days and for precipitation around 12 days. Of course, locally the impacts of land surface initialization can last much longer.
Second, we note the behavior of the red and blue curves in the first days of the forecast. For each variable except precipitation, the root-mean-square difference decreases during the first one to two weeks before increasing again. This reflects the influence of the identical atmosphere (and to some extent ocean) initialization on the soil moisture and temperature states and the surface fluxes. The ensemble members converge slightly during the period when there is still some coherence among the atmospheric states between ensemble members. After some time, this coherence has been lost because of the chaotic nature of the climate system, and the ensemble members begin to diverge again. The spread in soil wetness for maximum initial perturbation (red curve) does not fully reattain its high initial value until the third month of the simulations. The truly coupled nature of the land–atmosphere feedback system can be appreciated from this figure.
Despite the relative weaknesses in land–atmosphere coupling in CCSM4, which may mute the strength of the signal as a result of the land surface initialization in climate predictability, there is still ample evidence that the mechanisms and pathways of land–atmosphere coupling are present. If we take this as evidence that CCSM4 still gives useful but tempered estimates of land-driven predictability, we may assume for the moment that the relative changes in a changing climate are not greatly affected by this weakness in the mean behavior of the model. This is examined in the next section.
6. Changing predictability
To examine how predictability changes in CCSM4 as climate changes, we first show that the total variance of fields like near-surface air temperature is not constant from preindustrial to current to future conditions. Figure 12 shows the average change in the total variance of daily temperatures for JJA as the difference in the log of the variances. The ratio is taken so that in each case positive values suggest an increase in variance over time. Any colored areas are significant at the 95% confidence level.
There is, in fact, little change in total variance over land from preindustrial to current conditions. Most of the change in this model is over Arctic and tropical oceans (masked) associated with changes in sea ice and ENSO (DelSole et al. 2013b). When going from current climate to the RCP8.5 projection, a large fraction of continental areas show significant changes in total variance. Almost all the changes are positive, with the main concentration of negative changes over Greenland and the northernmost islands of Canada. Total temperature variance during JJA increases across most of the Northern Hemisphere tropics and subtropics (between the equator and about 30°N), much of Europe, the Arctic coasts of Asia and North America, and the Great Plains of the United States. Northeastern Brazil and the Kalahari of southern Africa show increases in the Southern Hemisphere.
The changes in total variance confound efforts to estimate the statistical significance of changes in predictability (cf. Guo et al. 2011). Nevertheless, we compare signal-to-total ratios as our metric for predictability using the same scales as in previous figures. Figure 13 shows the average change in signal-to-total ratios, again expressed as a difference in logs, where the signal comes from the realistic initialization of the land surface within each ensemble. Results are shown as averages over the second and third months for forecasts initialized at the start of May, June, and July.
From past to present, most of the changes are positive, implying greater predictability from the land surface. During this period there is not much change in total variance (Fig. 11), so the signal mainly increases at the expense of unforced variability (noise). The most pronounced areas of change are namely over those parts of South America and Africa where the behavior of C4 grassland phenology produces periodicity and, more interestingly, the central United States. This is an area that saw the greatest change in land cover from the mid-nineteenth century to the present—a vast expansion of agriculture that replaced more than 50% of the native vegetation over much of the region (Hurtt et al. 2011; Levis et al. 2012). The increase in predictability is most pronounced for the 1 June initialization. Most other regions of predictability change are small or transient; it would be difficult to attribute much consequence to them.
Small transient areas also dominate the change in predictability from current to future climate, but here there is a clear preference for negative changes. Positive values exist mainly over regions of the tropics and Southern Hemisphere where the RCP8.5 land use change scenarios prescribe advancing agriculture. Ironically, many of the decreases in the Southern Hemisphere and Sahel are also associated with expanding agriculture, over regions where coverage by C4 grasses is reduced. There is also some indication of continued increases in predictability over the central Great Plains, although spots of reduced predictability are also mixed in, especially for June initial conditions. But outside of these limited areas, there are many regions of reduced predictability on all continents. Many of these areas are coastal or in regions affected by oceanic variability; this model projects decreased future ENSO and tropical Atlantic predictability (Stevenson et al. 2012; DelSole et al. 2013b).
Finally, this experiment provides the opportunity to compare estimates of land–atmosphere coupling to predictability. Figure 14 shows the terrestrial coupling index of Dirmeyer (2011) calculated from the current climate simulation of CCSM4 (top) and the changes between climate regimes. This metric is the product of the standard deviation of daily latent heat flux and the correlation between latent heat flux (LHF) and root zone soil moisture (SM),
calculated for each of the months June, July, and August and then averaged together. Positive values indicate areas where soil moisture variations control surface fluxes; negative values are found where the fluxes drive soil moisture.
The terrestrial coupling index shows a pattern similar to estimates from other data sources (cf. Dirmeyer 2011) but generally weaker in magnitude. Most of the classical hot spots for this season are evident, including the Great Plains, eastern Europe, and the Indus valley. The Sahel is conspicuously absent. Negative values are found in the humid deep tropics, as well as in much of East Asia [a feature found also for other models by Dirmeyer (2011)].
Changes from preindustrial to current climate are generally small, highly regional, and not predominantly positive or negative. Projections for RCP8.5 show a global trend toward increased land–atmosphere coupling, consistent with the findings from other CMIP5 models (Dirmeyer et al. 2013).
Comparison between the terrestrial coupling index and predictability from land surface initialization (near-surface temperature signal-to-total variance ratio) reveals distinctly positive but weak spatial correlations across the boreal summer season. Values are generally in the range of 0.1–0.2 depending on the months and forecast lead times used. Thus, there does not appear to be a strong correspondence between land–atmosphere coupling and land-driven predictability in this model.
A large number of ensemble forecasts were generated from initial states taken from three separate 50-yr simulations with the NCAR CCSM4 climate model: one based on preindustrial climate conditions, one from a run branched from a historical climate run at the year 2000, and one based on the RCP8.5 future projection from 2095. The ensembles for each case were initialized from 15 of the 50 years based on an equal sampling of El Niño, La Niña, and neutral states calculated from each long simulation, with 15-member 3-month ensemble forecasts initialized on 1 May, 1 June, and 1 July. The ensembles are defined in three ways: identical initial land states but initial atmosphere and ocean chosen from each of the 15 years in the case (maximally perturbed); identical initial atmosphere and ocean states but land initial conditions from each of the 15 years; and identical initial atmosphere and ocean but small initial land perturbations taken from the days preceding the forecast start date. These runs form the basis of several studies into changing predictability in a changing climate—this study focuses on the role of the land surface.
Predictability in this study is defined from ensemble model statistics as the signal-to-total variance ratio and is confined to near-surface air temperature. The notion of the ensemble spread as a measure of noise cannot be verified in the real world because the real world only supplies us with a single realization of the time series of climate. We can only hope to measure directly the total variance of the real climate system—the partitioning between signal and noise must be inferred through the use of models, for instance by examining how changes in forecast initialization impact prediction skill (e.g., Koster et al. 2011).
Predictability owing to land surface initialization appears to be relatively weak in this model, compared to a variety of investigations with other climate models (e.g., Schlosser and Milly 2002; Koster et al. 2004, 2011), as is land–atmosphere coupling (e.g., Koster et al. 2006; Seneviratne et al. 2006; Dirmeyer 2011). Several potential causes of weak predictability are identified. These include a weaker than typical correlation between subsurface soil moisture and surface fluxes, insensitivity of the ABL to surface fluxes relative to entrainment and advection, and weak soil moisture memory (small lagged autocorrelation) compared to other model estimates and SCAN observations even after accounting for scale differences.
Some of the strongest predictability appears in subtropical regions with pronounced wet and dry seasons where C4 grasslands are prevalent and appear to be associated with a spurious oscillation in vegetation phenology during the dry season with a period of around 3 months. Otherwise, regions of relatively strong predictability are collocated with some areas of strong land–atmosphere feedback found in many other studies. When the impact of land surface initialization on predictability is quantified as the amplification of signal over that from atmosphere and ocean initialization (Fig. 9), a more coherent and widespread signal compared to Fig. 8 is revealed. This lends credence to the postulate that the land surface acts more as an amplifier or modulator of climate signals predominantly driven by SSTs (Koster et al. 2000; Reale and Dirmeyer 2002).
Changes in land-driven predictability from the mid-nineteenth to the late twenty-first century show clear associations with land use change, although the impacts are both positive and negative depending on the native vegetation being replaced. Generally speaking, expanding agriculture tends to enhance predictability because crops tend to respond more severely to climate anomalies than native vegetation.
Mechanistically, we expect regions of strong or changing feedbacks between land and atmosphere to correspond to regions of strong or changing climate predictability derived from the land state. This connection between coupling strength and predictability has been supposed as a potential means to improve seasonal forecasts (e.g., Seneviratne et al. 2006). CCSM4 projects an increase in the coupling between land and atmosphere in a warming climate, consistent with most other CMIP5 models (Dirmeyer et al. 2013). However, those changes do not always translate to increased predictability. There are more areas showing a decrease in land-driven predictability than an increase—most of the areas of increase are in the tropics. We do find some correspondence, but many other factors affect or modulate predictability.
These results are obviously model dependent. It has been long known that different models give different estimates of predictability on seasonal time scales (e.g., Shukla et al. 2000). While connections have been made between model systematic error and prediction skill (e.g., DelSole and Shukla 2010), the connection between error and predictability remains inferred. As with many recent modeling studies of land–climate interaction, a broader study including more models repeating the same investigation into predictability for different climate scenarios could elucidate a consensus response and quantify the degree of model uncertainty.
We thank David Lawrence for valuable discussion during the course of this analysis. The Climate Simulation Laboratory at NCAR's Computational and Information Systems Laboratory, sponsored by the National Science Foundation and other agencies, provided computing resources. This work was supported by joint funding of the Center for Ocean–Land–Atmosphere Studies (COLA) from the National Science Foundation (ATM-0830068), National Oceanic and Atmospheric Administration (NA09OAR4310058), and National Aeronautics and Space Administration (NNX09AN50G).